CUDASW++

CUDASW++ performs Smith-Waterman protein database searches by leveraging NVIDIA CUDA-enabled GPUs and multicore CPUs to accelerate sensitive local pairwise sequence alignment.

Key Features:

• Smith-Waterman algorithm: Implements the Smith-Waterman algorithm for maximally sensitive local pairwise protein sequence alignment.
• GPU acceleration: Executes alignment kernels on NVIDIA GPUs via the CUDA architecture (tested on Tesla GPUs and GeForce GTX 680/690 Kepler GPUs).
• CPU SIMD acceleration: Uses SSE-based vector execution units on CPUs to accelerate alignment computations.
• CUDA PTX SIMD instructions: Employs CUDA PTX SIMD video instructions to extend data parallelism beyond the conventional SIMT model on GPUs.
• CPU–GPU workload distribution: Automatically distributes workload between CPUs and GPUs based on their respective computational strengths to optimize throughput.
• Performance benchmarks: Demonstrates speed-ups up to 2.9× on single‑GPU GeForce GTX 680 and 3.2× on dual‑GPU GeForce GTX 690 versus CUDASW++ 2.0, reaching peak performance of 119.0 and 185.6 GCUPS respectively.
• Comparative performance: Reportedly outperforms other sequence alignment tools such as SWIPE and BLAST+ in benchmark comparisons.
• Implementation and target architecture: Implemented in CUDA C++ and PTX assembly and targets Kepler‑architecture GPUs, combining CPU and GPU SIMD instructions for concurrent execution.
• Database evaluation: Evaluated using the Swiss-Prot protein database.

Scientific Applications:

• Protein database searches: High-sensitivity protein sequence database searches using Smith-Waterman alignments (e.g., Swiss-Prot).
• Pairwise local alignment: Sensitive pairwise local sequence alignment of protein sequences.
• High-throughput database scanning: Accelerated processing of large-scale protein sequence databases where quadratic-time Smith-Waterman computation is a bottleneck.

Methodology:

Implements the Smith-Waterman algorithm with SSE-based CPU vectorization and GPU acceleration using CUDA PTX SIMD instructions, automatically partitions workload between CPUs and GPUs, and is implemented in CUDA C++ and PTX assembly targeting Kepler GPUs.